I've seen confliction on if *bsd or linux is better, this (hopefully) isn't that surprising to anyone.
You should do a PPS throughput analysis of your own to see which OS works better on the hardware that you plan to use. Drivers, and the susceptibility of the kernel to livelock, are where there may be differences in performance.
Finally, it appears as if, contrary to what the articles are saying, not many people are actively considering such a move. However, it is more common in smaller businesses starting new locations or building out.
DEC's gateway to the Internet ran on host-based routers - DEC Alphas running Digital UNIX with turbochannel FDDI cards - from 1994 to sometime in 1999-ish (I stopped being responsible for it in 1998). I started with a pair and had suffered one all-night upgrade to eight when the PPS load of some AltaVista announcement pushed the pair over the edge into livelock.
What about better case situations?* IE:
toe cards
TOE won't help you, you aren't terminating TCP sessions on the box. At least you shouldn't be. Don't let anyone talk you into also running a web server.
custom kernel
This could be useful, if the kernel is able to handle all packet forwarding in the interrupt or polling input service routine.
no moving parts (ie: hard drive, maybe fans if possible)
That'll certainly help with reliability, as well as dual power supplies.
up-to-date software packages with internal coders to fix ugly bugs, etc actual research into what packages & hardware would be best
Both of those things, or a support agreement from one of the vendors that's trying to make the host-based open-source router business model work. Stephen